Ocfs2 users - Oct 2004 - poor 2 node performance

Hi,

I've been setting ocfs2 up on a two node 'cluster', using gnbd
(network block device) to talk to a shared disk on a third node.

build/install was very straightforward and I have the file system
mounted on both nodes just fine. However, I've been rather
disappointed by the performance: With just one node actively
using the fs (but with the other mounted) I'm finding that file
system operations are running *very* slowly -- a linux kernel
build takes 5 times longer, with the system blocked on IO most of
the time. 

Looking at the node that is supposedly quiescent, I see that
kernel thread "events/0" is burning 100% of the CPU. Sure enough,
syslog is filling up rapidly with messages of the form:

Oct 17 22:01:13 breakout-0 kernel: process_vote: type: MODIFY, lockid:
2751426560, action: (11) <NULL>, num_ident: 1, alive: 1, write: 0,
change2719924224, node=1, seqnum=551559, response=1

I guess I could disable the logging, but, I'm still rather
surprised about the amount of communication between the nodes.
Is this normal?  I guess I was hoping that ocfs2 used some kind
of hierarchical read/write/existence locking on fs subtrees.

Performance with just a single node mounted is perfectly
respectable.

Any suggestions?

Thanks,
Ian

Hi Ian

most of the multinode stuff is very much in flux right now (kurt is
basically rewriting it).  killing syslogd would definitely help ;) but
it won't be the silver bullet. anyay yeah multinode just right now
sucks, but the new dlm/net/nm stuff will be plugged in soon

Wim

On Sun, Oct 17, 2004 at 10:11:43PM +0100, Ian Pratt
wrote:
> 
> Hi,
> 
> I've been setting ocfs2 up on a two node 'cluster', using gnbd
> (network block device) to talk to a shared disk on a third node.
> 
> build/install was very straightforward and I have the file system
> mounted on both nodes just fine. However, I've been rather
> disappointed by the performance: With just one node actively
> using the fs (but with the other mounted) I'm finding that file
> system operations are running *very* slowly -- a linux kernel
> build takes 5 times longer, with the system blocked on IO most of
> the time. 
> 
> Looking at the node that is supposedly quiescent, I see that
> kernel thread "events/0" is burning 100% of the CPU. Sure enough,
> syslog is filling up rapidly with messages of the form:
> 
> Oct 17 22:01:13 breakout-0 kernel: process_vote: type: MODIFY, lockid:
2751426560, action: (11) <NULL>, num_ident: 1, alive: 1, write: 0,
change2719924224, node=1, seqnum=551559, response=1
> 
> I guess I could disable the logging, but, I'm still rather
> surprised about the amount of communication between the nodes.
> Is this normal?  I guess I was hoping that ocfs2 used some kind
> of hierarchical read/write/existence locking on fs subtrees.
> 
> Performance with just a single node mounted is perfectly
> respectable.
> 
> Any suggestions?
> 
> Thanks,
> Ian
> 
> _______________________________________________
> Ocfs2-users mailing list
> Ocfs2-users@oss.oracle.com
> http://oss.oracle.com/mailman/listinfo/ocfs2-users

On Sun, Oct 17, 2004 at 10:11:43PM +0100, Ian Pratt
wrote:
> I guess I could disable the logging, but, I'm still rather
> surprised about the amount of communication between the nodes.
> Is this normal?  I guess I was hoping that ocfs2 used some kind
> of hierarchical read/write/existence locking on fs subtrees.
	This is normal.  There is no fancy hierarchicalness.

Joel

-- 

"Not everything that can be counted counts, and not everything
 that counts can be counted."
        - Albert Einstein 

Joel Becker
Senior Member of Technical Staff
Oracle Corporation
E-mail: joel.becker@oracle.com
Phone: (650) 506-8127